与PPL组合的SIMD对齐问题

Question

我试图将数组的元素与SIMD并行求和。为了避免锁定，我使用可组合的线程局部，它并不总是在16字节上对齐因为_mm_add_epi32抛出异常

concurrency::combinable<__m128i> sum_combine;
int length = 40; // multiple of 8
concurrency::parallel_for(0, length , 8, [&](int it)
{
    __m128i v1 = _mm_load_si128(reinterpret_cast<__m128i*>(input_arr + it));
    __m128i v2 = _mm_load_si128(reinterpret_cast<__m128i*>(input_arr + it + sizeof(uint32_t)));
    auto temp = _mm_add_epi32(v1, v2);
    auto &sum = sum_combine.local();   // here is the problem 

    TRACE(L"%dn", it);
    TRACE(L"add %xn", &sum);
    ASSERT(((unsigned long)&sum & 15) == 0);
    sum = _mm_add_epi32(temp, sum);
}
);

这里是可组合的定义，来自ppl.h

template<typename _Ty>
class combinable
{
private:
// Disable warning C4324: structure was padded due to __declspec(align())
// This padding is expected and necessary.
#pragma warning(push)
#pragma warning(disable: 4324)
    __declspec(align(64))
    struct _Node
    {
        unsigned long _M_key;
        _Ty _M_value;                   // this might not be aligned on 16 bytes
        _Node* _M_chain;
        _Node(unsigned long _Key, _Ty _InitialValue)
            : _M_key(_Key), _M_value(_InitialValue), _M_chain(NULL)
        {
        }
    };

有时对齐是可以的，代码工作得很好，但大多数时候它不工作

我已经尝试使用下面的命令，但是这不能编译

union combine 
{
        unsigned short x[sizeof(__m128i) / sizeof(unsigned int)];
        __m128i y;
};
concurrency::combinable<combine> sum_combine;
then auto &sum = sum_combine.local().y; 

有什么建议纠正对齐问题，仍然使用组合式。

在x64上，它工作得很好，因为默认的16字节对齐。在x86上有时存在对齐问题。

Answer 1

使用unaligned load加载sum

auto &sum = sum_combine.local();

#if !defined(_M_X64) 
if (((unsigned long)&sum & 15) != 0)
{
    // just for breakpoint means, sum  is unaligned.
    int a = 5;
}
auto sum_temp = _mm_loadu_si128(&sum);
sum = _mm_add_epi32(temp, sum_temp);
#else
sum = _mm_add_epi32(temp, sum);
#endif

Answer 2

由于_mm_add_epi32使用的sum变量未对齐，您需要使用未对齐的加载/存储(_mm_loadu_si128/_mm_storeu_si128)显式加载/存储sum。变化:

sum = _mm_add_epi32(temp, sum);

:

__m128i v2 = _mm_loadu_si128((__m128i *)&sum);
v2 = _mm_add_epi32(v2, temp);
_mm_storeu_si128((__m128i *)&sum, v2);